Triplet longitudinal masked autoencoder for predicting individualized functional connectome development during...

Area of Science:

• Neuroscience
• Developmental Neuroscience
• Medical Imaging

Background:

• Resting-state functional MRI (rs-fMRI) is key for infant brain functional connectivity (FC) studies.
• Predicting infant FC trajectories from incomplete longitudinal data is vital for understanding brain development and identifying disorders.
• Current deep learning methods struggle with temporal inconsistencies and missing data in infant FC prediction.

Purpose of the Study:

• To develop a novel method for accurately predicting the full dynamic developmental trajectory of infant brain functional connectivity.
• To address challenges posed by scarce longitudinal data and irregular missing scans in infant neuroimaging.
• To improve the temporal consistency and accuracy of FC predictions during infancy.

Main Methods:

• Proposed a Triplet Longitudinal Masked Autoencoder (TL-MAE) for dynamic infant FC trajectory prediction.
• Implemented a longitudinally consistent prediction strategy for robust FC generation.
• Utilized an FC-specific Masked Autoencoder pre-trained on large datasets.
• Developed a dual triplet network with an identity conditional module for individualized age-based predictions.

Main Results:

• The TL-MAE method demonstrated more accurate and temporally consistent predictions of FC developmental trajectories.
• The model successfully captured individualized features in infant brain development.
• Experimental results on 696 longitudinal infant fMRI scans validated the method's superiority over state-of-the-art techniques.

Conclusions:

• The TL-MAE offers a significant advancement in predicting dynamic infant brain functional connectivity.
• This method enhances understanding of normal and abnormal neurodevelopmental trajectories.
• TL-MAE provides a robust tool for individualized prediction of infant brain development from rs-fMRI data.